McGill algorithm for precipitation nowcasting by lagrangian extrapolation (MAPLE) applied to the South Korean radar network. Part II: Real-time verification for the summer season

Abstract

The MAPLE system has been implemented in real-time in Korea since June 2008, producing forecasts up to 6 hours every 10 minutes. An object-oriented verification method has been applied for the summer season (June-July-August) over the Korean Peninsula to evaluate and understand the characteristics of the forecast results. The CRA (contiguous rain area) approach is used to decompose the total error into the different error components; location, pattern, and volume errors. The mean displacement error is smaller than 20 km up to the 3-h forecasts and increases with forecast time. The ratio between the displacement (location) error and the total error is less than 7% even for a 3-h forecast. This result indicates that MAPLE produces reliable forecast in terms of precipitation location. However, the pattern error is larger than 90% of the total error. Contingency scores that are defined with different categories of rain intensity and displacement errors show the outstanding performance up to 2.5 hours. MAPLE overpredicts rain areas with the threshold of 1 mm h −1 rain intensity throughout forecast periods. However, the heavy rainfall events are poorly predicted due to the inherent limitation of extrapolation-based nowcasting technique.

Computer Vision in Precipitation Nowcasting: Applying Image Quality Assessment Metrics for Training Deep Neural Networks

“…Differs from ConvLSTM and ConvGRU, which directly use convolution for the state transitions, TrajGRU uses convolution to generate flow fields of two steps and calculates new states by a warping procedure. We noticed that the operation of TrajGRU is quite similar to three steps of the widely applied numerical McGill algorithm [11][12][13][14]: radar echo tracking, radar reflectivity advecting (by a semi-Lagrangian extrapolation) and error evaluating. However, while this type of traditional method needs predefined spatial filters and translation vector to detect the motions, deep learning methods can automatically learn features from data in an end-to-end fashion [5].…”

“…Another advantage of deep neural networks is that they can be easily applied for various spatial scales, image sizes and image resolutions, simply by stacking more layers or changing the convolutional kernel sizes. Especially, Shi et al [6] showed that modifying the training loss function appropriately can increase the accuracy in predicting heavy rainfall events, which is a drawback of McGill models [14].…”

Optimize Short-Term Rainfall Forecast with Combination of Ensemble Precipitation Nowcasts by Lagrangian Extrapolation

“…This result seems to be based on the fact that the ensemble forecast was made by considering the uncertainty of rainfall propagation, i.e., by considering more feasible cases of rainfall propagation. Also, this limitation, which is the quality of three hour forecast is far lower than that of one hour and two hours can be found in the previous studies in Korea [108][109][110].…”

Improving Radar Echo Lagrangian Extrapolation Nowcasting by Blending Numerical Model Wind Information: Statistical Performance of 16 Typhoon Cases

“…Since the average mountain height in Taiwan is around 2 km with peaks up to 4 km, it is a natural environment for evaluating MAPLE's performance over complex terrain. In previous studies, MAPLE has shown its capability for nowcasting up to 2-6 h depending on the regions being implemented Zawadzki 2002, 2004;Turner et al 2004;Bellon et al 2010;Lee et al 2010, Mandapaka et al 2012. MAPLE was first applied and configured over Taiwan in 2018 (Pan et al 2018).…”